An Efficient Inference Framework for Early-exit Large Language Models

TL;DR

This paper introduces an efficient inference framework tailored for early-exit large language models, addressing unique challenges and achieving up to 1.25x speedup over traditional methods.

Contribution

It proposes novel batch inference and KV cache management techniques specifically designed for early-exit LLMs, filling a gap in existing inference frameworks.

Findings

Achieves up to 1.25x speedup compared to full-layer inference.

Introduces batch processing until all sequences meet confidence thresholds.

Develops KV cache management method for early-exit models.

Abstract

Building efficient inference framework has gained increasing interests for research community. Early-exit models, a variant of LLMs, improves the inference efficiency of LLMs by skipping rest layers and directly generate output tokens when they are confident enough. However, there is no work of LLM inference framework that takes early-exit models into consideration. This is non-trivial as prior art on LLM inference cannot be directly applied to early-exit models. In this work, we solves two key challenges in building efficient inference framework for early-exit models: (1) batch inference at iteration-level granularity; and (2) KV cache management. For the former, we propose to process the batch until all sequences surpass the early-exit confidence threshold. For the latter, we propose to fill the KV cache of rest layers before the iteration terminates. Our evaluation shows that, compared with the original vLLM operating at full layers, our solution achieves up to 1.25x speed up.